We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: MaGNET: Uniform Sampling from Deep Generative Network Manifolds Without Retraining

Abstract: Deep Generative Networks (DGNs) are extensively employed in Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and their variants to approximate the data manifold, and data distribution on that manifold. However, training samples are often obtained based on preferences, costs, or convenience producing artifacts in the empirical data distribution e.g., the large fraction of smiling faces in the CelebA dataset or the large fraction of dark-haired individuals in FFHQ. These inconsistencies will be reproduced when sampling from the trained DGN, which has far-reaching potential implications for fairness, data augmentation, anomaly detection, domain adaptation, and beyond. In response, we develop a differential geometry based sampler -- coined MaGNET -- that, given any trained DGN, produces samples that are uniformly distributed on the learned manifold. We prove theoretically and empirically that our technique produces a uniform distribution on the manifold regardless of the training set distribution. We perform a range of experiments on various datasets and DGNs. One of them considers the state-of-the-art StyleGAN2 trained on FFHQ dataset, where uniform sampling via MaGNET increases distribution precision and recall by 4.1% & 3.0% and decreases gender bias by 41.2%, without requiring labels or retraining.
Comments: 13 pages, 14 pages Appendix, 23 figures
Subjects: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV)
Cite as: arXiv:2110.08009 [cs.LG]
  (or arXiv:2110.08009v2 [cs.LG] for this version)

Submission history

From: Ahmed Imtiaz Humayun [view email]
[v1] Fri, 15 Oct 2021 11:12:56 GMT (43932kb,D)
[v2] Mon, 18 Oct 2021 01:41:01 GMT (43938kb,D)

Link back to: arXiv, form interface, contact.